Sir, the support for distributed locks is only added because some customers wish to use the file system catalog in object storage. If I do not support the demands of these users, then the entire file system catalog does not need to rely on distributed locks at all. What I would like to understand is, is it because you believe that the file system catalog must rely on distributed locks that you have ceased further support for it? If that is the case, I can directly remove all parts of the file system catalog that use distributed locks.
---- Replied Message ---- | From | Daniel Weeks<dwe...@apache.org> | | Date | 07/15/2024 00:08 | | To | dev@iceberg.apache.org | | Cc | | | Subject | Re: Re: Core:support redis and http lock-manager | Iisoda, Unfortunately, I don't agree with your assessment. The problems with file system based catalog implementations are inherent and steps taken to address them are not adequate to have confidence in the implementation. Commit atomicity is not solved as it relies on locking, which has a number of fundamental issues that will continue to plague these implementations. Locking distributes the complexity to clients who all need to participate properly, but issues like lost locks, lock timeouts, clock skew and lock sequencing/deadlocks (for more complicated commit scenarios) are all introduced as new issues. These are examples of problems that exist in Hive locking today. I don't think we can call this solved. Beyond that, there are a whole set of operations like CTAS, RTAS, IF [NOT] EXISTS, DROP, RENAME, etc. that are not atomic with this model. These are all important functions used in data warehouses and the more you try to solve these, the more you end up relying on external systems to track table state and metadata, which ends up not being a file system based catalog. Adding complexity to file system catalog implementations by introducing more dependencies to try to address these issues just confuses users by giving the impression that this is a legitimate alternative to a real catalog implementation. I don't think we should be adding functionality here and we should probably deprecate or relocate the file system catalog to the test codebase like the in-memory catalog. -Dan On Sun, Jul 14, 2024 at 4:12 AM lisoda <lis...@yeah.net> wrote: At present, the file system based catalogues have the following problems (this is what I can think of at the moment, perhaps not comprehensive). 1. does not support renaming operations 2. commit does not support atomicity 3. atomic delete (sorry I don't understand why we need it?What scenarios need this?)? 4.Some features are missing, such as: support for views. We have solved issue 2, and issue 1 is tolerated by users for the time being. Regarding issue 3, we don't know what the specific problem is that it relates to. From what I've seen so far, I think most of the problems can actually be solved.Filesystem-based catalogues can achieve at least a read-commit level of isolation.For OLAP systems, it is sufficient for most scenarios. I'm not sure if I'm being overly optimistic about this issue, so please let me know if there's anything wrong with my opinion.Thank you. 在 2024-07-13 02:07:42，"Ryan Blue" <b...@databricks.com.INVALID> 写道： I think one of the main questions is whether we want to support locking strategies moving forward. These were needed in early catalogs that didn't have support for atomic operations (HadoopCatalog and GlueCatalog). Now, Glue supports atomic commits and we have been discouraging the use of HadoopCatalog, which is a purely filesystem-based implementation for a long time. One thing to consider is that external locking does solve a few of the challenges of the filesystem-based approach, but doesn't help with many of the shortcomings of the HadoopCatalog, like being able to atomically delete or rename a table. (Operations that are very commonly used in data engineering!) Maybe we should consider moving Hadoop* classes into a separate iceberg-hadoop module, along with the LockManager to make it work somewhat better. Personally, I'd prefer deprecating HadoopCatalog and HadoopTableOperations because of their serious limitations. But moving these into a separate module seems like a good compromise. That would also avoid needing to add dependencies to core, like Redis for lock implementations. Ryan On Thu, Jul 11, 2024 at 10:42 PM lisoda <lis...@yeah.net> wrote: Currently, the only lockManager implementation in iceberg-core is InMemoryLockManager. This PR extends two LockManager implementations, one based on the Redis, and another based on the Rest-API. In general, most users use redisLockManager is sufficient to cope with most of the scenarios, for redis can not meet the user's requirements, we can let the user to provide a RestApi service to achieve this function. I believe that, for a long time, these two lock-manager's will satisfy most of the customer's needs. If someone could review this PR, that would be great. PR: https://github.com/apache/iceberg/pull/10688 SLACK: https://apache-iceberg.slack.com/archives/C03LG1D563F/p1720761992982729 -- Ryan Blue Databricks
